Large CNC machine tools are characterized by long guide rail travel and are often subjected to complex environmental disturbances in industrial settings. Existing volumetric error detection methods generally suffer from limited accuracy and low efficiency. To address the above shortcomings, a measurement method for volume error of large machine tools based on beam drift compensation and model optimization is proposed. To mitigate accuracy degradation caused by beam drift in long-distance measurements, a differential compensation method for angular drift is proposed based on the difference in beam polarization states, which is the common transmission of reference light and measurement light and beam splitting detection. At the same time, the effect of the non-parallelism of the two beams on the roll angle measurement was analyzed, and a high-parallelism two-beam generation module was constructed by utilizing the characteristics of retroreflector. On this basis, a five-degree-of-freedom (5-DOF) error online measurement system was developed, which achieved high-precision and high-efficiency acquisition of the geometric error of the machine-sheet axis. In order to further improve the accuracy and applicability of the volume error model of large machine tools, an optimized volume error model suitable for various three-axis machine tools is constructed. Starting from the homogeneous transformation model, the influence mechanism of Abbe error and Bryan error caused by the non-collinearity between the error measurement axis and the motion axis is introduced into the model. Performance tests of the measurement system and compensation experiments for machine tool volume error measurements were carried out in the laboratory and industrial sites respectively. The results show that within the 3 m measurement range, the standard deviation of the angular error of the 5-DOF measurement system is below 0.5″, and the standard deviation of the straightness error is less than 0.6 μm. After error measurement and compensation, the diagonal positioning error of the machine tool is reduced by 51.6%. This method can achieve precise online measurement and active compensation of volume errors of large machine tools, and has good industrial application prospects.